Regensburg 2025 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

BP: Fachverband Biologische Physik

BP 29: Focus Session: Innovations in Research Software Engineering (joint session BP/DY)

BP 29.4: Vortrag

Donnerstag, 20. März 2025, 16:00–16:15, H44

Invert pattern forming systems with BayesFlow to bridge the gap from simulation to experimental observation — •Hans Olischläger — Interdisciplinary Center for Scientific Computing (IWR) — Heidelberg University

The description of experimental systems by complex spatial models, be it with (stochastic) partial differential equations, agent-based simulation or otherwise, is often the condensation of all the central scientific hypotheses regarding a particular object of study.

I argue, that making progress in this kind of modelling is currently hindered by the lack of a tool that enables solving the following inverse problem: Given an observation, determine all the model configurations that are able to produce it. In other words, what is the posterior probability of all model configurations given some (set of) experimental data.

Instead of just preaching that in theory a Bayesian treatment would be nice, I will then continue to present such a tool: amortized Bayesian inference (as implemented in the software package BayesFlow). I will give examples on the classical Gierer-Meinhardt pattern forming PDE and a biophysical model, the Min system, which is used by E. coli to control cell division.

I will also take a step back to give a broader picture of the newly available statistical methods that support complex spatial modelling and their limitations. The aim is to provide some guidance on what you can and cannot infer from your state-of-the-art scientific simulator given observations, and how to do it.

Keywords: simulation-based inference; pattern formation; inverse problem; Python; Julia